Recent work on Generative Adversarial Networks for honeypot generation, most notably the HoneyGAN Pots framework, has shown that decoys produced by a learned generator can evade existing honeypot detectors more reliably than hand-crafted ones. That result is encouraging, but it is only a first step. A trained generator still produces decoys that are static after training: it does not learn from how attackers actually behave once the decoys are deployed, it provides no tamper-proof record of what it generated, and it assumes a class of compute resources that does not exist on industrial edge nodes. This paper is a position paper that takes the future-work agenda outlined by Gabrys et al. and develops it into a concrete research design. We propose an integrated framework, HoneyGAN+, that couples a conditional GAN to a reinforcement-learning policy trained on observed attacker engagement, anchors every generated decoy on a permissioned blockchain to give responders a verifiable audit trail, and uses knowledge distillation together with federated learning to push inference onto resource-constrained Industry 5.0 endpoints. We further argue that individual decoys should be replaced by coordinated multi-agent ecosystems whose internal references are mutually consistent, and we propose a standardized evaluation pipeline (detection resistance, engagement, longevity, and intelligence value) to make claims comparable across studies. The paper also discusses dual-use risks and a governance model that, in our view, ought to be developed before broad deployment rather than after. We are explicit throughout: the contribution is a research design with enough technical detail to support implementation, not a fully built and benchmarked system.